Role of Earth Observations for Crop Area Estimates in ... · acquire images with high geometric...

G44

1PROCEEDINGS ICAS VII Seventh International Conference on Agricultural Statistics I Rome 24-26 October 2016

Role of Earth Observations for Crop Area Estimates in

Africa. Experiences from the AGRICAB Project

Michele Downie Consorzio-ITA

Via Farnesi 25/A 55100 Lucca, Italy

[email protected]

David Remotti Consorzio-ITA

Via Farnesi 25/A 55100 Lucca, Italy

[email protected]

Tomaso Ceccarelli * Alterra, Wageningen UR

Droevendaalsesteeg, 6708 PB

Wageningen, the Netherlands

[email protected]

DOI: 10.1481/icasVII.2016.g44b

https://urlsand.esvalabs.com/?u=https%3A%2F%2Fdoi.org%2F10.1481%2FICASVII.2016.G44B&e=1b20e90c&h=60dd718e&f=n&p=y

G44


ABSTRACT

Despite major technological advancements, Earth Observations (EO) are nowadays

seldom used in national agricultural statistical systems at an operational level. This is with the

exception of a few countries and is especially true in Africa (but for Morocco, South Africa

and a few other cases). Moving from small tests confined to research to operational contexts,

poses a number of challenges. There are many aspects that still restrict the use of this

technology: the high cost of some of the images, the level of expertise, overall organisation

and institutional arrangements needed to operationally run the different components of the

system.The FP7 AGRICAB project provided an important opportunity to evaluate the role of

EO specifically for crop area estimates using spatial frames in Africa. This is with reference

to: 1) the construction and stratification of a (point) sampling frame, 2) the support to the

ground surveys, and 3) the use of the image classifications to further improve the accuracyof

the estimates from the surveys alone. The approach was tested in selected areas in

Mozambique, Senegal and Kenya. The contribution of EO for the last aspect cited was

evaluated only for the latter two areas, where high geometric resolution images (RapidEye)

were used. Sampling frames were built based on high geometric resolution images accessible

through Google Earth. This contributed significatively to the frame construction and to

increased sampling efficiency through land cover stratification. As to the use of EO for

survey preparation/execution, the images also provided invaluable support to the geolocation

of sampling units especially if combined with GPS. Finally, EO data were used for reducing

the variance of ground estimates. If a high correlation between the classification of the image

and the ground truth exists, it is possible to produce estimates with a lower sampling error.

From a theoretical point of view a perfect correlation between spectral signatures of a generic

crop and the corresponding parcel, would promote pure remote sensing approaches such as

pixel counting and sampling frames combined withground surveys unnecessary.

Unfortunately this high correlation rarely exists and in most cases satellite images cannot be

used directly. Confusion matrixes generated from the supervised classifications of the test

areas showed a very low overall accuracy and hence, a very low contribution to the reduction

of the variance of the estimates which was evaluated further in terms of net efficiency and

cost efficiency. This was to be expected, considering the characteristics of most Sub-Saharan

agricultural landscapes, i.e.: small fields vs. pixel size, continuous and mixed cropping, low

planting densities. The use of multi-temporal images and the combination of optical and radar

EO data can probably increase such correlation although a rigorous cost/benefits analysis

would be needed to evaluate its added value also in view of new satellite products (e.g. those

provided by the Sentinel missions).

Keywords: spatial frames, satellites, relative efficiency

G44


1. Introduction

There are two main methods to derive crop area statistics from Earth Observations (EO)

data: pure remote sensing approaches such as pixel counting, and methods combining field

survey data and image classification results (FAO 2015). Pixel counting is the more direct

way, although it is often criticized for the bias which can be introduced (Gallego at al., 2008

and Gallego et al., 2010). This seems to be especially true with reference to smallholders

farming systems in Sub-Saharan Africa.

Despite major technological advancements, EO are nowadays seldom used in national

agricultural statistical systems in an operational way. There are a limited number of countries

where this occurs worldwide and, apart from Morocco, South Africa and a few other cases,

this is especially true for Africa.

Moving from small tests confined to research to operational contexts, poses a number of

challenges. There are many aspects that still restrict the use of this technology: depending on

their use, the high cost of some of the images, the level of expertise required, overall

organisation and institutional arrangements needed to run operationally the different

components of the system.

The FP7 AGRICAB project provided an important opportunity to evaluatethe

contribution of EO specifically for crop acreage estimates using spatial frames in Africa. This

is with reference to three levels in the methodology: 1) construction and stratification of a

point sampling frame, 2) support to the ground survey, and 3) its use to further improve the

accuracy of the estimates from the survey alone.

The project initially targeted selected areas in Senegal, Kenya and Mozambique and

together with national partners mandated with agricultural statistics in the countries. EO

products were effectively used in all areas. However only in the case of Senegal and Kenya,

where images with high geometric resolution were available, it was possible to cover all cited

levels of application.Due to the cloud cover, for MozambiqueEO products were used only

atthe first two levels.

2. Materials and Methods

2.1 Materials

The areas selected were the District (Département) of Nioro Du Rip in Senegal, and the

County of Kakamega-Butere in Kenya. As mentioned, only for these areas it was possible to

acquire images with high geometric resolution.In Mozambique the area selected was the

District (Distrito) of Inharrime. The three areas are indicated in Figure 1.

G44


Figure 1: The three areas of interest

The EO images used belong to the RapidEye satellite constellation. For the cited study

areas the provider programmed the image acquisition at a convenient time with reference to

the presence of the targeted crops. Harvesting periods relating to the two area tests were

indicated by experts from the respective Ministries of Agriculture who have knowledge of the

climate and the characteristics of the selected test areas. In Kakamega-Butere the period was

15 May - 30 June while in Nioro Du Rip, 15 August - 30 September.

G44


Table 1: Main features of RapidEye images.

As it discussed in the next sections, satellite images are complementing survey data. A

description of the content of such information as well as the way in wich ground data are

collected is also given in the next setions.

2.2 Sampling Frame Construction

The sampling frame is the most important element in methodology developed. Its

functions are:

to enumerate all the units of the population;

to label and stratify the same units based on a limited number of land cover

classes;

to allow the extraction of sampling units for a specific statistical survey;

to subsequently extrapolate to the universe the values derived from the sample.

The construction of the frame requires the following steps:

Defining the units of the population. These units are represented by geographic points

located at the vertices of a grid of 500 x 500m. All points within an administrative boundary

represent the population whose parameters (crop area in this case) we intend to estimate.

G44


Building the frame.This implies assigning (labelling) the land cover type to each unit

based on very few and simple classes. This is done through on screen visual interpretation of

satellite imagery. The images do not necessarily need to be of the same year of the survey, nor

need to be taken during the growing season; the higher the resolution, the better, since it

allows a more accurate identification of the named classes. In this respect very high resolution

(VHR), i.e. sub-metric satellite images freely available on Google Earth or similar sources

were found to be the most suitable basis to derive the land cover types related to the points

making the population.Only in a few cases high resolution images were not available.

Stratification. The land cover classes assigned to the points in the frame provide a basis

for its further stratification.

Satellite images are importantfor the construction of the frame. Without their

contribution the said frame would rely on ortho-photographs in digital format (which are

simply not available in most African contexts) or be built using only topographic maps. These

maps in most cases do not contain information useful for the stratification hence reducing the

efficiency of the statistical system as a whole.

Figure 2: Labelling of the point

2.3 Support to ground-survey

The ground survey consists of the identification of crop types and other additional

information in each sampling point. In this way objective, verifiable and repeatable data on

crop occurrence can be collected based on rigorous sampling schemes. The ground

observations are carried out by the surveyor whit reference to an area (usually a circle of fixed

radius) around each point.

When carrying out the ground survey, the surveyor is faced with a number of challenges

such as poor road network and lack of topographic information. Areas are also often

G44


largelyuninhabited,with long distances between points, and may present poor accessibility as

well as specific dangers. Therefore the survey must be carefully planned beforehand and

constantly revised on the ground..

Recent, very high resolution EO products are required for the purpose.There are usually

three types of maps available to the surveyors, all having satellite information as a backdrop

image. Examples are given in Figure 3. An “indexmap” is used as a topographic base map to

locate the tiles. A “tile map” is a portion of the base map, generally at a 1:50.000 scale, and

contains the points which should be visited by a surveyor. A “(sample frame) point

map”,generally at a 1:2.000 scale, which is used together with a hand held GPS to identify

and access the sampling point. Such map with the sampling units superimposed on satellite

images can greatly facilitate the proper recognition of their position. For the geolocation of

the points one should not rely exclusively on most GPS used in these types of surveys, which

have known limitations in their geometric accuracy.

It is important to underline that the type of land cover and crop type information which

needs to be associated to each point, cannot be derived directly from the satellite images used

in the construction of the frame and in its stratification. This is due to the fact that such

images are usually taken before the ground survey (i.e. when the cloud cover is minimal).

Therefore they usually do not allow for a proper identification of the crop type or, if this is the

case, may represent crops which have changed over the different cropping seasons.

Figure 3: Field Maps (from bottom left: Index, Tile and Point Map)

G44


Using the point maps, the surveyor can also plan the best itinerary to reach the different

sampling units. Each surveyor has access to a series of field tools (the same field maps, field

forms and a GPS). These would allow him a) to reach, as quickly as possible, each sampling

unit and to identify with high accuracy (from a geographical point of view)the point of

observation of the same unit, b) to identifythe type of crop according to a pre-defined legend

and classification rules. An illustration of the field tools is given in figure 4.

Figure 4: Survey field tools

Based on the experience in the different surveys conducted we can conclude that VHR

images, combined with other baseline information and the location of the same sampling units

provide an invaluable support for survey planning and execution. They make it possible to

carry out the survey quickly and rigorously without the need of contacting the farmers.

2.4 Use of the image classifications to further improve the accuracy of the

estimates from the ground surveys

This is, at least in principle, one of the most important aspects related to the application

of EO data for crop area estimates. If a high correlation between the classified images and the

ground truth exists, it is possible to generate crop area estimates characterized by a lower

sampling error (expressed in terms of coefficient of variation, or CV). The RapidEye images

acquired for the two study areas have been processed with the aim to verify the degree of the

correlation between the spectral signatures of the main crop types and the corresponding

G44


ground truth observed during the survey. The following procedures have been utilized for

processing the satellite images:

Image pre-processing.includes geo-referencing through registration of images,

mosaicing of the RapidEye tiles with the same dates and same atmospheric conditions,

production of masks to remove non-agricultural areas from the classification (otherwise

introducing a bias).

Spectral signature extraction.The final outcomes of this step area number of Regions

Of Interest (ROI) which are then used to guide the classification and to provide the ground

truth locations needed for generating the confusion matrix.

The extraction was carried out using the observations collected during the surveys. Data

varyaccording to the study areas and the respective methodologies. For Kenya the ground

truth was based on the "Field Sample Points" grid(around 1.000 points) and on an additional

sample of "Photo-clusters"(each cluster is a meshof 100 points with 50 meters spacing). This

sample was classifiedby visual interpretation of aerial ortho-photos collected by the

Department of Resource Surveys and Remote Sensing (DRSRS) during the survey period. In

Senegal the available ground truth came from the "Field Sample Points" grid data (again

around 1.000 points collected in the field) and the "Parcel Segments"data collected by the

“Direction de l'Analyse, de la Prevision et des StatistiquesAgricoles” (DAPSA) of the

Ministry of Agriculture, with the support of the “Centre de SuiviEcologique” (CSE). Crop

sample areaswere collectedby GPS during the “Enquêteagricole 2013” in conjunction with the

AGRICAB point frame survey. The procedure needs to be adjusted to each area due to

differences in the type of the ground truth data i.e. to differences in the classification rules

applied and tothe geometry of the data (points or areas). Such information is often collected in

difficult environmental conditions and this goes sometimes to the detriment of its quality and

accuracy. Moreover important factors related to the spectral signature (phenological state,

health conditions, farmingpractices, water stress, etc.) could not always be captured during the

surveys.

Supervised classification of the images. The final mask, the mosaics and the ROI

dataset were imported in a GIS environment. In Kenya the classification was carried out

based on the QGIS and GRASS open source solutions, while in Senegal the software ERDAS

was used.The ROI dataset was used to generate the signature files, one for each mosaic. The

classification was performed applying the maximum likelihood classifier (pixel based).

Considering the available information it was decided to classify the images focusing on

the most representative crops: Sugarcane and Maize for Kenya and Millet, Maize, Sorghum

and Groundnuts for Senegal. All the other land use classes were grouped as "other land

cover/land use” (LC/LU).

The following tasks were performed for the classification:

use ofsub-classes in order to cover as much as possible the various spectral signatures that

characterize each LC/LU class;

adding the most important natural/non-agriculture sub-classes to reduce the variance

in the classification of the agricultural classes;

aggregating the sub-classes corresponding to the various spectral classes obtained from

the images in LC/LU classes and all the non-agricultural sub-classes as "Other LC/LU";

filtering the classified pixels with a confidence level of at least 66% to70%;

calculating the confusion matrix and the accuracies;

G44


An example of the original image and of the classification is given in Figure 5.

Figure 5: Example of the original image and the classification in Senegal, Nioro Du

Rip. From top: composite RGB 543 and classified raster image

The accuracies derived from the confusion matrix should not be regarded as a post

classification assessment, but rather as an indicator of the validity of the classificationin

further reducing the estimates based on ground survey data alone.

Second stratification.The classification provides information on land cover and crop

types that could be used to achieve better estimates of crop areas. The principle is that this

G44


information, known for the whole of the population and not only for the sampling units, can

be used to build a stratification of the population in homogeneous strata where the variance of

the estimates is expected to be significantly lower than in the population as a whole.

A new stratification is performed on top of the initial one (see 2.2), using a binary

classification for each point in the sample frame.For instance in the case of maize four new

strata are derived as a combination of the original strata, and the points classified as maize (or

“not-maize”) in the image.

3. Results

As shown in the tables 2 and 3 below, the confusion matrices generated in the two

different study areas show a very low overall accuracy in the classification: around 45% for

Senegal and around 40% for Kenya. For the main crops occurring in each area omission and

commission errors are relatively better.

Table 2: Confusion matrix Nioro Du Rip, Senegal

G44


Table 3: Confusion matrix Kagamega-Butere (Kenya)

Possible reasons behind these low accuracies can be ascribed to several factors. Some

factors are of general applicability while others are especially important in the agricultural

landscapes targeted in this study (i.e. characterized by small-holder farmers, with low levels

of management and crop densities). Altogether they all contribute to the fact that the

relationship between crops on the ground and their spectral signature is less evident:

Geometric resolution: related to the pixels size. In this case parts of the cropped areas

can be characterised by fields which are too small to be shaped by the pixels. This aspect

could, at least in principle, be reduced by the introduction of EO data with higher geometric

resolution.

Atmospheric conditions: different atmospheric conditions can occur, even at local

level, when the dates of the images are distant in time.

Crop phenology: each crop speciesand even variety develops according to different

phenological phasesto which representative spectral signatures can be associated.It is thus

very difficult that crops reach the most representative spectral signature at same time.

Seeding date: even within a species and variety having a specific phenology, the

starting (i.e. seeding) date of the crop cycle can change according to several factors including

climate and, locally, topography and soil conditions. Management choices of farmers are also

very important.

Continuous cropping: the agro-ecologic conditions allow some overlap in the growing

cycles of crops especially in sub-tropical environments.

Intercropping: in the same parcel different crops are sown at the same time or with a

little time lag (relay cropping) resulting in a non-distinctive spectral signature, being a mix

between two existing classes. A similar problem is given in case of weeds.

Crop density: it relates to farmers management practises, soil fertility, pathologies,

water stress. Soil in the background alters the natural spectral signature.

G44


Difficulty in determining the best date to collect the satellite images: it is due to the

unpredictability of the weather conditions and the lack of synchronization of cultural

practices/crops phenology.

The area estimations can be now calculated based on the binary stratification described

in section 2. The crops where the classification performs better are selected for this purpose,

i.e. Groundnuts in Senegaland Maize in Kenya (Butere-Mumias only). Results are given in

Table 4 and 5:

Table 4: Estimations with and without the image classification: Groundnuts, Senegal

Approach Crop area (ha) StdDeviation CV

only survey 72.229.70 2.713.10 3.76

survey + classification 72.597.02 2.688.12 3.7

Table 5: Estimations with and without the image classification: Maize, Kenya

Comparing the results above the gain in precision (expressed in terms of reduction of

the CV) achieved introducing the classification in the area estimates can be observed. Since

there is only a minor decrease in the CV for both areas and crops, one can conclude that the

contribution of the image classifications to further improve the accuracy of the estimates from

the surveyis altogether very modest.

A more formal way to assess the contribution of EO data can be expressed in terms of

Relative Efficiency (RE), which is the ratio between the variance of the ground survey area

estimate and the variance after this estimate has been corrected withthe aid of classified

satellite images. RE values close to 1 indicate that the EO contribution towards reducing the

sampling variance estimations is very low:

RE = (Var.withouteo / Var.with eo)

Where:

Var.without eo = Variance estimation obtained without the contribution of EO data

Var.with eo= Variance estimation with the contribution of the EO data.

In both areas the RE is in the order of 1.01 (1.009 for Senegal and 1.007 for Kenya).If

the cost components are known, also the cost efficiency can be computed. This is expressed

in terms of Net Relative Efficiency (NRE) and calculated as follows:

Approach Crop area (ha) StdDeviation CV

only survey 29.537.73 2.633.56 8.92

survey + classification 29.421.20 2.615.51 8.89

G44


NRE = RE x (Costs without eo / Costs with eo)

Where:

Costs without eo = Costs without the contribution of EO data (i.e. ground survey costs).

Costs with eo = Costs considering the contribution of EO data (i.e. ground survey costs, cost

of the images, of the radiometric and geometric corrections and of the classification).

In the two selected areas the NRE is in the order of 0.21, which indicates a rather low

performance in terms of cost-efficiency.

4. Conclusions

As far as the generation of crop area estimates the contribution of EO was evaluated for

two of the areas (and the main crops cultivated within) selected as use cases in the project

AGRICAB. As to the contribution to the third level of application (image classifications to

improve accuracy of estimates from the ground surveys) the results were evaluated in terms of

RE and cost efficiency. The result indicates a low contribution in improving the estimates

and, as a consequence and due to the type of images used, an even lower cost-efficiency.

Nevertheless, the other two levels of application, i.e. the construction and stratification

of the sampling frame, and the support to ground surveys, were deemed very important,

although it was not possible to further quantify their contribution.

In the future, in order to increase the efficiency in the use of EO data, a multi

spectral/temporal approach can be further explored. A minimum of at least two images must

be foreseen. A critical issue related to the areas of interest and similar environments is the

presence of cloud cover during the main growing seasons. This poses a constraint on the

proposed multi temporal approach using optical satellites only. In perspective, a combination

of SAR (Sentinel 1) and optical satellites (Sentinel 2) may provide a useful opportunity for

this type of applications.

REFERENCES

Carfagna, E. (2013) Evaluating the cost-efficiency of remote sensing in developing

countries.FAO Global Strategy, first SAC meeting, Rome.

Ceccarelli, T. and Remotti, D. (2012)Service utility report: agricultural survey optimisation

service in Malawi.ESA report. Paris.

Ceccarelli, T., Remotti, D., Haub, C. (2013)Agricultural survey optimization service: Malawi

report on activities, 2013. ASO Malawi final meeting, Lilongwe.

FAO (2015)Technical Report onCost – Effectiveness of Remote Sensing for Agricultural

Statistics in Developing and EmergingEconomies. Technical Report Series GO-09-2015,

FAO, Rome.

G44


Gallego J.F., Craig M., Michaelsen J., Bossyns B., Fritz S. (Editors), (2008) Best practices for

crop area estimation with Remote Sensing. GEOSS Community of Practice Ag 0703a, JRC

Scientific and Technical Reports.

Gallego, F.J., Carfagna, E. and Baruth, B.,(2010)Accuracy, objectivity and efficiency of

remote sensing for agriculture statistics. in: R. Benedetti, M. Bee, G. Espa& F. Piersimoni

(eds.) Agricultural Survey Methods, pp. 193–211. New York, John Wiley and Sons.

Giovacchini A., Remotti D., Monaci L., Downie M, Imala V., Mosquito Patricio F., Tauacale

F., Diop B., (2015)AGRICAB AGRICULTURAL STATISTICS – TECHNICAL

REPORTDeliverable 31.3.

Date post:	19-Aug-2020
Category:	Documents
Upload:	others
View:	0 times
Download:	0 times

Role of Earth Observations for Crop Area Estimates in ... · acquire images with high geometric...

Documents